!nvidia-smi
Tue Aug 9 12:52:02 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.32.03 Driver Version: 460.32.03 CUDA Version: 11.2 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla P100-PCIE... Off | 00000000:00:04.0 Off | 0 |
| N/A 36C P0 26W / 250W | 0MiB / 16280MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
from google.colab import drive
drive.mount('/content/gdrive')
Mounted at /content/gdrive
%cd /content/gdrive/MyDrive/Atelier_Google_Colab/Projet_8
# List files to make sure we're in the expected directory.
!ls
/content/gdrive/MyDrive/Atelier_Google_Colab/Projet_8 data_redim P08_01_notebook_modelisation.ipynb saved src
#!pip uninstall imgaug
#!pip install git+https://github.com/aleju/imgaug.git
Found existing installation: imgaug 0.4.0
Uninstalling imgaug-0.4.0:
Would remove:
/usr/local/lib/python3.7/dist-packages/imgaug-0.4.0.dist-info/*
/usr/local/lib/python3.7/dist-packages/imgaug/*
Proceed (y/n)? y
Successfully uninstalled imgaug-0.4.0
Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/
Collecting git+https://github.com/aleju/imgaug.git
Cloning https://github.com/aleju/imgaug.git to /tmp/pip-req-build-vq9g2i1k
Running command git clone -q https://github.com/aleju/imgaug.git /tmp/pip-req-build-vq9g2i1k
Requirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (1.15.0)
Requirement already satisfied: numpy>=1.15 in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (1.21.6)
Requirement already satisfied: scipy in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (1.7.3)
Requirement already satisfied: Pillow in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (7.1.2)
Requirement already satisfied: matplotlib in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (3.2.2)
Requirement already satisfied: scikit-image>=0.14.2 in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (0.18.3)
Requirement already satisfied: opencv-python in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (4.6.0.66)
Requirement already satisfied: Shapely in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (1.8.2)
Requirement already satisfied: imageio in /usr/local/lib/python3.7/dist-packages (from imgaug==0.4.0) (2.9.0)
Requirement already satisfied: tifffile>=2019.7.26 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.14.2->imgaug==0.4.0) (2021.11.2)
Requirement already satisfied: networkx>=2.0 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.14.2->imgaug==0.4.0) (2.6.3)
Requirement already satisfied: PyWavelets>=1.1.1 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.14.2->imgaug==0.4.0) (1.3.0)
Requirement already satisfied: kiwisolver>=1.0.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib->imgaug==0.4.0) (1.4.4)
Requirement already satisfied: cycler>=0.10 in /usr/local/lib/python3.7/dist-packages (from matplotlib->imgaug==0.4.0) (0.11.0)
Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib->imgaug==0.4.0) (3.0.9)
Requirement already satisfied: python-dateutil>=2.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib->imgaug==0.4.0) (2.8.2)
Requirement already satisfied: typing-extensions in /usr/local/lib/python3.7/dist-packages (from kiwisolver>=1.0.1->matplotlib->imgaug==0.4.0) (4.1.1)
Building wheels for collected packages: imgaug
Building wheel for imgaug (setup.py) ... done
Created wheel for imgaug: filename=imgaug-0.4.0-py3-none-any.whl size=971122 sha256=06bb79d041875cd32e55c4cc94a443f33613719bf73654539fe2919a3bc92e0b
Stored in directory: /tmp/pip-ephem-wheel-cache-o3p89m3x/wheels/0c/78/b5/9303fae9d5e03df1f319adfe4e6534180b5c3232de11bc9a2f
Successfully built imgaug
Installing collected packages: imgaug
Successfully installed imgaug-0.4.0
%reload_ext autoreload
%autoreload 2
%matplotlib inline
import os
import random
import cv2
import time
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.python.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.metrics import MeanIoU, OneHotMeanIoU
from keras.models import load_model
import imgaug as ia
import imgaug.augmenters as iaa
# Personal libraries
import src.datagenerator as dg
import src.models as models
import src.helpers as helpers
# Images
leftImg8bit_path = 'data_redim/leftImg8bit/'
# Masques
gtFine_path = 'data_redim/gtFine/'
checkpoint_path = 'saved/checkpoint/'
models_path = 'saved/models/'
history_path = 'saved/history/'
if not os.path.isdir('saved'):
os.makedirs(checkpoint_path)
os.makedirs(models_path)
os.makedirs(history_path)
DIM_X = 256
DIM_Y = 256
INPUT_SHAPE = (DIM_X,DIM_Y,3)
NB_CLASSES = 8
monitor_val = 'val_mean_iou'
learning_rate = 1e-5
nbr_epochs = 50
patience = 14
# Recupère les chemins d'accès des fichiers dans le dossier précisé
def getPathFiles(dir_path, pattern = ""):
train_path = dir_path + 'train/'
test_path = dir_path + 'test/'
val_path = dir_path + 'val/'
train_path_files = []
test_path_files = []
val_path_files = []
# Train set
for dir in os.listdir(train_path):
for file in os.listdir(train_path + dir):
if pattern in file:
train_path_files.append(train_path + dir + '/' +file)
# Test set
for dir in os.listdir(test_path):
for file in os.listdir(test_path + dir):
if pattern in file:
test_path_files.append(test_path + dir + '/' +file)
# Val set
for dir in os.listdir(val_path):
for file in os.listdir(val_path + dir):
if pattern in file:
val_path_files.append(val_path + dir + '/' +file)
return train_path_files, test_path_files, val_path_files
# Listes des chemins des images
x_train_path, x_test_path, x_val_path = getPathFiles(leftImg8bit_path)
# Listes des chemins des masques
y_train_path, y_test_path, y_val_path = getPathFiles(gtFine_path, "labelIds.png")
# Echantillon préparé pour les tests
X_test, y_test = helpers.get_data_prepared(x_test_path[0:10], y_test_path[0:10], (DIM_X,DIM_Y))
data_redim_path = 'data_redim'
list_dir = os.listdir()
if data_redim_path not in list_dir:
os.mkdir('data_redim')
for dir_1 in os.listdir('data'):
if dir_1 != 'img':
os.mkdir(data_redim_path + '/' + dir_1)
for dir_2 in os.listdir('data' + '/' + dir_1):
os.mkdir(data_redim_path + '/' + dir_1 + '/' + dir_2)
for dir_3 in os.listdir('data' + '/' + dir_1 + '/' + dir_2):
os.mkdir(data_redim_path + '/' + dir_1 + '/' + dir_2 + '/' + dir_3)
all_images_path = x_train_path + x_test_path + x_val_path + y_train_path + y_test_path + y_val_path
for path in all_images_path:
img = cv2.resize(cv2.imread(path), (1024, 512))
new_path = path.replace('data', 'data_redim')
cv2.imwrite(new_path, img)
else:
print(f'Dossier {data_redim_path} déjà créé')
Dossier data_redim déjà créé
X = np.array([cv2.resize(cv2.imread(path_X), (DIM_X, DIM_Y)) for path_X in x_train_path[0:5]])
y = np.array([cv2.resize(helpers.convert_mask(cv2.imread(path_y,0)), (DIM_X, DIM_Y)) for path_y in y_train_path[0:5]])
X_augmented, y_augmented = helpers.augment_data(X, y)
y_augmented_arg = np.argmax(y_augmented, axis=3)
for i in range(len(X)):
plt.subplots(1, 4, figsize=(21,5))
plt.subplot(1, 4, 1)
plt.imshow(X[i])
plt.subplot(1, 4, 2)
plt.imshow(X_augmented[i])
plt.subplot(1, 4, 3)
plt.imshow(y_augmented_arg[i])
plt.subplot(1, 4, 4)
plt.imshow(X_augmented[i])
plt.imshow(y_augmented_arg[i], alpha=0.3)
plt.show()
data_train = dg.DataGenerator(x_train_path, y_train_path, batch_size=20, dim=(DIM_X, DIM_Y), augmentation = False)
data_val = dg.DataGenerator(x_val_path, y_val_path, batch_size=20, dim=(DIM_X, DIM_Y), augmentation = False)
data_train_augmented = dg.DataGenerator(x_train_path, y_train_path, batch_size=10, dim=(DIM_X, DIM_Y), augmentation = True)
data_val_augmented = dg.DataGenerator(x_val_path, y_val_path, batch_size=10, dim=(DIM_X, DIM_Y), augmentation = True)
import json
synthese = {}
train_times = {}
if os.path.exists(models_path + 'train_times.json'):
with open(models_path + 'train_times.json', 'r') as f:
train_times = json.load(f)
def train_model(model, model_name, loss_function, augmented=False):
filepath_check = checkpoint_path + model_name + '/'
model_path = models_path + model_name
history_path_file = history_path + model_name + '_history.json'
if os.path.exists(model_path):
if loss_function == "categorical_crossentropy":
model = load_model(model_path)
else:
model = load_model(model_path, custom_objects={'dice_loss': dice_loss,'total_loss': total_loss}, compile=True)
print(f'Récupération du modèle {model.name}')
training_time = train_times[model_name]
with open(history_path_file, 'r') as f:
history = json.load(f)
hist_df = pd.DataFrame.from_dict(history)
else:
start_train = time.time()
# compile
model.compile(optimizer=Adam(learning_rate=learning_rate), loss=loss_function, metrics=[OneHotMeanIoU(num_classes=NB_CLASSES, name='mean_iou')])
mc = ModelCheckpoint(mode='max', filepath=filepath_check, monitor=monitor_val, save_best_only='True', save_weights_only='True')
es = EarlyStopping(mode='max', monitor=monitor_val, patience=patience, verbose=1)
callbacks = [mc, es]
# fit
if augmented:
history = model.fit(data_train_augmented, epochs=nbr_epochs, validation_data=data_val_augmented, use_multiprocessing=True, workers=2, callbacks=callbacks)
else:
history = model.fit(data_train, epochs=nbr_epochs, validation_data=data_val, use_multiprocessing=True, workers=2, callbacks=callbacks)
# convert the history.history dict to a pandas DataFrame:
hist_df = pd.DataFrame(history.history)
# save history
with open(history_path_file, mode='w') as f:
hist_df.to_json(f)
# Save the model
model.load_weights(filepath_check)
model.save(model_path)
end_train = time.time()
training_time = end_train - start_train
train_times[model_name] = training_time
with open(models_path + 'train_times.json', 'w') as f:
json.dump(train_times, f)
return model, hist_df, training_time
def predict_model(model, model_name, loss_type, training_time, show_pred=False):
loss_score, mean_iou_score = model.evaluate(data_val)
print("Pour le meilleur modèle on obtient :")
print("mean_iou :", mean_iou_score)
print("loss :", loss_score)
start_pred = time.time()
y_pred=model.predict(X_test)
y_pred_argmax=np.argmax(y_pred, axis=3)
end_pred = time.time()
predict_time = end_pred - start_pred
if show_pred:
for index in range(10):
print(index)
plt.imshow(X_test[index])
plt.imshow(y_pred_argmax[index],alpha = 0.4)
plt.show()
synthese[model_name] = helpers.add_model_in_synthese(loss_type, loss_score, mean_iou_score, training_time, predict_time)
return y_pred_argmax
# Build model
model = models.build_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_base'
model.summary()
softmax
Model: "U-Net"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 256, 256, 3 0 []
)]
conv2d (Conv2D) (None, 256, 256, 64 1792 ['input_1[0][0]']
)
batch_normalization (BatchNorm (None, 256, 256, 64 256 ['conv2d[0][0]']
alization) )
activation (Activation) (None, 256, 256, 64 0 ['batch_normalization[0][0]']
)
conv2d_1 (Conv2D) (None, 256, 256, 64 36928 ['activation[0][0]']
)
batch_normalization_1 (BatchNo (None, 256, 256, 64 256 ['conv2d_1[0][0]']
rmalization) )
activation_1 (Activation) (None, 256, 256, 64 0 ['batch_normalization_1[0][0]']
)
max_pooling2d (MaxPooling2D) (None, 128, 128, 64 0 ['activation_1[0][0]']
)
conv2d_2 (Conv2D) (None, 128, 128, 12 73856 ['max_pooling2d[0][0]']
8)
batch_normalization_2 (BatchNo (None, 128, 128, 12 512 ['conv2d_2[0][0]']
rmalization) 8)
activation_2 (Activation) (None, 128, 128, 12 0 ['batch_normalization_2[0][0]']
8)
conv2d_3 (Conv2D) (None, 128, 128, 12 147584 ['activation_2[0][0]']
8)
batch_normalization_3 (BatchNo (None, 128, 128, 12 512 ['conv2d_3[0][0]']
rmalization) 8)
activation_3 (Activation) (None, 128, 128, 12 0 ['batch_normalization_3[0][0]']
8)
max_pooling2d_1 (MaxPooling2D) (None, 64, 64, 128) 0 ['activation_3[0][0]']
conv2d_4 (Conv2D) (None, 64, 64, 256) 295168 ['max_pooling2d_1[0][0]']
batch_normalization_4 (BatchNo (None, 64, 64, 256) 1024 ['conv2d_4[0][0]']
rmalization)
activation_4 (Activation) (None, 64, 64, 256) 0 ['batch_normalization_4[0][0]']
conv2d_5 (Conv2D) (None, 64, 64, 256) 590080 ['activation_4[0][0]']
batch_normalization_5 (BatchNo (None, 64, 64, 256) 1024 ['conv2d_5[0][0]']
rmalization)
activation_5 (Activation) (None, 64, 64, 256) 0 ['batch_normalization_5[0][0]']
max_pooling2d_2 (MaxPooling2D) (None, 32, 32, 256) 0 ['activation_5[0][0]']
conv2d_6 (Conv2D) (None, 32, 32, 512) 1180160 ['max_pooling2d_2[0][0]']
batch_normalization_6 (BatchNo (None, 32, 32, 512) 2048 ['conv2d_6[0][0]']
rmalization)
activation_6 (Activation) (None, 32, 32, 512) 0 ['batch_normalization_6[0][0]']
conv2d_7 (Conv2D) (None, 32, 32, 512) 2359808 ['activation_6[0][0]']
batch_normalization_7 (BatchNo (None, 32, 32, 512) 2048 ['conv2d_7[0][0]']
rmalization)
activation_7 (Activation) (None, 32, 32, 512) 0 ['batch_normalization_7[0][0]']
max_pooling2d_3 (MaxPooling2D) (None, 16, 16, 512) 0 ['activation_7[0][0]']
conv2d_8 (Conv2D) (None, 16, 16, 1024 4719616 ['max_pooling2d_3[0][0]']
)
batch_normalization_8 (BatchNo (None, 16, 16, 1024 4096 ['conv2d_8[0][0]']
rmalization) )
activation_8 (Activation) (None, 16, 16, 1024 0 ['batch_normalization_8[0][0]']
)
conv2d_9 (Conv2D) (None, 16, 16, 1024 9438208 ['activation_8[0][0]']
)
batch_normalization_9 (BatchNo (None, 16, 16, 1024 4096 ['conv2d_9[0][0]']
rmalization) )
activation_9 (Activation) (None, 16, 16, 1024 0 ['batch_normalization_9[0][0]']
)
conv2d_transpose (Conv2DTransp (None, 32, 32, 512) 2097664 ['activation_9[0][0]']
ose)
concatenate (Concatenate) (None, 32, 32, 1024 0 ['conv2d_transpose[0][0]',
) 'activation_7[0][0]']
conv2d_10 (Conv2D) (None, 32, 32, 512) 4719104 ['concatenate[0][0]']
batch_normalization_10 (BatchN (None, 32, 32, 512) 2048 ['conv2d_10[0][0]']
ormalization)
activation_10 (Activation) (None, 32, 32, 512) 0 ['batch_normalization_10[0][0]']
conv2d_11 (Conv2D) (None, 32, 32, 512) 2359808 ['activation_10[0][0]']
batch_normalization_11 (BatchN (None, 32, 32, 512) 2048 ['conv2d_11[0][0]']
ormalization)
activation_11 (Activation) (None, 32, 32, 512) 0 ['batch_normalization_11[0][0]']
conv2d_transpose_1 (Conv2DTran (None, 64, 64, 256) 524544 ['activation_11[0][0]']
spose)
concatenate_1 (Concatenate) (None, 64, 64, 512) 0 ['conv2d_transpose_1[0][0]',
'activation_5[0][0]']
conv2d_12 (Conv2D) (None, 64, 64, 256) 1179904 ['concatenate_1[0][0]']
batch_normalization_12 (BatchN (None, 64, 64, 256) 1024 ['conv2d_12[0][0]']
ormalization)
activation_12 (Activation) (None, 64, 64, 256) 0 ['batch_normalization_12[0][0]']
conv2d_13 (Conv2D) (None, 64, 64, 256) 590080 ['activation_12[0][0]']
batch_normalization_13 (BatchN (None, 64, 64, 256) 1024 ['conv2d_13[0][0]']
ormalization)
activation_13 (Activation) (None, 64, 64, 256) 0 ['batch_normalization_13[0][0]']
conv2d_transpose_2 (Conv2DTran (None, 128, 128, 12 131200 ['activation_13[0][0]']
spose) 8)
concatenate_2 (Concatenate) (None, 128, 128, 25 0 ['conv2d_transpose_2[0][0]',
6) 'activation_3[0][0]']
conv2d_14 (Conv2D) (None, 128, 128, 12 295040 ['concatenate_2[0][0]']
8)
batch_normalization_14 (BatchN (None, 128, 128, 12 512 ['conv2d_14[0][0]']
ormalization) 8)
activation_14 (Activation) (None, 128, 128, 12 0 ['batch_normalization_14[0][0]']
8)
conv2d_15 (Conv2D) (None, 128, 128, 12 147584 ['activation_14[0][0]']
8)
batch_normalization_15 (BatchN (None, 128, 128, 12 512 ['conv2d_15[0][0]']
ormalization) 8)
activation_15 (Activation) (None, 128, 128, 12 0 ['batch_normalization_15[0][0]']
8)
conv2d_transpose_3 (Conv2DTran (None, 256, 256, 64 32832 ['activation_15[0][0]']
spose) )
concatenate_3 (Concatenate) (None, 256, 256, 12 0 ['conv2d_transpose_3[0][0]',
8) 'activation_1[0][0]']
conv2d_16 (Conv2D) (None, 256, 256, 64 73792 ['concatenate_3[0][0]']
)
batch_normalization_16 (BatchN (None, 256, 256, 64 256 ['conv2d_16[0][0]']
ormalization) )
activation_16 (Activation) (None, 256, 256, 64 0 ['batch_normalization_16[0][0]']
)
conv2d_17 (Conv2D) (None, 256, 256, 64 36928 ['activation_16[0][0]']
)
batch_normalization_17 (BatchN (None, 256, 256, 64 256 ['conv2d_17[0][0]']
ormalization) )
activation_17 (Activation) (None, 256, 256, 64 0 ['batch_normalization_17[0][0]']
)
conv2d_18 (Conv2D) (None, 256, 256, 8) 520 ['activation_17[0][0]']
==================================================================================================
Total params: 31,055,752
Trainable params: 31,043,976
Non-trainable params: 11,776
__________________________________________________________________________________________________
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, "categorical_crossentropy", augmented=False)
Modèle : U-Net_base Récupération du modèle U-Net
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "categorical_crossentropy", training_time, show_pred=True)
25/25 [==============================] - 449s 18s/step - loss: 1.0450 - mean_iou: 0.3021 Pour le meilleur modèle on obtient : mean_iou : 0.30205994844436646 loss : 1.0450489521026611 0
1
2
3
4
5
6
7
8
9
from tensorflow.python.keras.losses import categorical_crossentropy
from tensorflow.python.keras import backend as K
def dice_coeff(y_true, y_pred):
smooth = 1.
y_true_f = K.cast(K.flatten(y_true), K.floatx())
y_pred_f = K.cast(K.flatten(y_pred), K.floatx())
intersection = K.sum(y_true_f * y_pred_f)
score = (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
return score
def dice_loss(y_true, y_pred):
loss = 1 - dice_coeff(y_true, y_pred)
return loss
def total_loss(y_true, y_pred):
loss = categorical_crossentropy(y_true, y_pred) + (3*dice_loss(y_true, y_pred))
return loss
# Build model
model = models.build_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_dice_loss'
softmax
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, dice_loss, augmented=False)
Modèle : U-Net_dice_loss Récupération du modèle U-Net
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "dice_loss", training_time, show_pred=True)
25/25 [==============================] - 38s 1s/step - loss: 0.3764 - mean_iou: 0.2936 Pour le meilleur modèle on obtient : mean_iou : 0.29356127977371216 loss : 0.37635427713394165 0
1
2
3
4
5
6
7
8
9
# Build model
model = models.build_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_combine_loss'
softmax
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, total_loss, augmented=False)
Modèle : U-Net_combine_loss Récupération du modèle U-Net
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "combine_loss", training_time, show_pred=True)
25/25 [==============================] - 38s 1s/step - loss: 2.4823 - mean_iou: 0.3026 Pour le meilleur modèle on obtient : mean_iou : 0.3026266098022461 loss : 2.482293128967285 0
1
2
3
4
5
6
7
8
9
# Build model
model = models.build_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_augmented'
softmax
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, total_loss, augmented=True)
Modèle : U-Net_augmented Récupération du modèle U-Net
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "combine_loss", training_time, show_pred=True)
25/25 [==============================] - 39s 2s/step - loss: 2.3302 - mean_iou: 0.3037 Pour le meilleur modèle on obtient : mean_iou : 0.30368149280548096 loss : 2.330190658569336 0
1
2
3
4
5
6
7
8
9
# Build model
model = models.build_resnet50_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_augmented'
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/resnet/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5 94773248/94765736 [==============================] - 1s 0us/step 94781440/94765736 [==============================] - 1s 0us/step
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, total_loss, augmented=True)
Modèle : ResNet50_U-Net_augmented Récupération du modèle ResNet50_U-Net
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "combine_loss", training_time, show_pred=True)
25/25 [==============================] - 38s 1s/step - loss: 2.3114 - mean_iou: 0.3132 Pour le meilleur modèle on obtient : mean_iou : 0.31320345401763916 loss : 2.311370849609375 WARNING:tensorflow:5 out of the last 5 calls to <function Model.make_predict_function.<locals>.predict_function at 0x7fb2f3e88c20> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details. 0
1
2
3
4
5
6
7
8
9
# Build model
model = models.build_vgg16_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_base'
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, total_loss, augmented=False)
Modèle : VGG16_U-Net_base Récupération du modèle VGG16_U-Net
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "combine_loss", training_time, show_pred=True)
25/25 [==============================] - 33s 1s/step - loss: 2.3707 - mean_iou: 0.3053 Pour le meilleur modèle on obtient : mean_iou : 0.30532243847846985 loss : 2.3706603050231934 0
1
2
3
4
5
6
7
8
9
# Build model
model = models.build_vgg16_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_augmented'
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg16/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5 58892288/58889256 [==============================] - 0s 0us/step 58900480/58889256 [==============================] - 0s 0us/step
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, total_loss, augmented=True)
Modèle : VGG16_U-Net_augmented Epoch 1/50 297/297 [==============================] - 1606s 5s/step - loss: 3.4115 - mean_iou: 0.1375 - val_loss: 3.3118 - val_mean_iou: 0.1576 Epoch 2/50 297/297 [==============================] - 253s 849ms/step - loss: 2.7443 - mean_iou: 0.1908 - val_loss: 2.8461 - val_mean_iou: 0.1831 Epoch 3/50 297/297 [==============================] - 250s 839ms/step - loss: 2.6220 - mean_iou: 0.2095 - val_loss: 2.7006 - val_mean_iou: 0.1903 Epoch 4/50 297/297 [==============================] - 252s 846ms/step - loss: 2.5556 - mean_iou: 0.2163 - val_loss: 2.6380 - val_mean_iou: 0.2051 Epoch 5/50 297/297 [==============================] - 253s 843ms/step - loss: 2.5222 - mean_iou: 0.2244 - val_loss: 2.6038 - val_mean_iou: 0.2141 Epoch 6/50 297/297 [==============================] - 249s 834ms/step - loss: 2.4930 - mean_iou: 0.2300 - val_loss: 2.5923 - val_mean_iou: 0.2114 Epoch 7/50 297/297 [==============================] - 250s 838ms/step - loss: 2.4847 - mean_iou: 0.2318 - val_loss: 2.5554 - val_mean_iou: 0.2223 Epoch 8/50 297/297 [==============================] - 252s 843ms/step - loss: 2.4647 - mean_iou: 0.2371 - val_loss: 2.5528 - val_mean_iou: 0.2244 Epoch 9/50 297/297 [==============================] - 248s 830ms/step - loss: 2.4517 - mean_iou: 0.2423 - val_loss: 2.5291 - val_mean_iou: 0.2376 Epoch 10/50 297/297 [==============================] - 250s 836ms/step - loss: 2.4374 - mean_iou: 0.2481 - val_loss: 2.5417 - val_mean_iou: 0.2328 Epoch 11/50 297/297 [==============================] - 251s 840ms/step - loss: 2.4219 - mean_iou: 0.2517 - val_loss: 2.5477 - val_mean_iou: 0.2330 Epoch 12/50 297/297 [==============================] - 255s 850ms/step - loss: 2.4012 - mean_iou: 0.2551 - val_loss: 2.5222 - val_mean_iou: 0.2388 Epoch 13/50 297/297 [==============================] - 253s 843ms/step - loss: 2.3894 - mean_iou: 0.2573 - val_loss: 2.5285 - val_mean_iou: 0.2419 Epoch 14/50 297/297 [==============================] - 249s 835ms/step - loss: 2.3731 - mean_iou: 0.2593 - val_loss: 2.5134 - val_mean_iou: 0.2345 Epoch 15/50 297/297 [==============================] - 250s 839ms/step - loss: 2.3663 - mean_iou: 0.2616 - val_loss: 2.4945 - val_mean_iou: 0.2466 Epoch 16/50 297/297 [==============================] - 250s 837ms/step - loss: 2.3523 - mean_iou: 0.2639 - val_loss: 2.4783 - val_mean_iou: 0.2472 Epoch 17/50 297/297 [==============================] - 250s 836ms/step - loss: 2.3376 - mean_iou: 0.2667 - val_loss: 2.4736 - val_mean_iou: 0.2472 Epoch 18/50 297/297 [==============================] - 249s 833ms/step - loss: 2.3298 - mean_iou: 0.2686 - val_loss: 2.4727 - val_mean_iou: 0.2442 Epoch 19/50 297/297 [==============================] - 247s 828ms/step - loss: 2.3251 - mean_iou: 0.2687 - val_loss: 2.4665 - val_mean_iou: 0.2463 Epoch 20/50 297/297 [==============================] - 248s 829ms/step - loss: 2.3141 - mean_iou: 0.2706 - val_loss: 2.4778 - val_mean_iou: 0.2415 Epoch 21/50 297/297 [==============================] - 251s 840ms/step - loss: 2.3046 - mean_iou: 0.2731 - val_loss: 2.4792 - val_mean_iou: 0.2474 Epoch 22/50 297/297 [==============================] - 252s 842ms/step - loss: 2.2976 - mean_iou: 0.2740 - val_loss: 2.4628 - val_mean_iou: 0.2498 Epoch 23/50 297/297 [==============================] - 245s 821ms/step - loss: 2.2872 - mean_iou: 0.2762 - val_loss: 2.4822 - val_mean_iou: 0.2431 Epoch 24/50 297/297 [==============================] - 248s 832ms/step - loss: 2.2833 - mean_iou: 0.2757 - val_loss: 2.4635 - val_mean_iou: 0.2468 Epoch 25/50 297/297 [==============================] - 249s 834ms/step - loss: 2.2744 - mean_iou: 0.2777 - val_loss: 2.4555 - val_mean_iou: 0.2486 Epoch 26/50 297/297 [==============================] - 248s 831ms/step - loss: 2.2680 - mean_iou: 0.2789 - val_loss: 2.4813 - val_mean_iou: 0.2397 Epoch 27/50 297/297 [==============================] - 250s 838ms/step - loss: 2.2611 - mean_iou: 0.2806 - val_loss: 2.4770 - val_mean_iou: 0.2358 Epoch 28/50 297/297 [==============================] - 250s 837ms/step - loss: 2.2592 - mean_iou: 0.2813 - val_loss: 2.4702 - val_mean_iou: 0.2414 Epoch 29/50 297/297 [==============================] - 247s 828ms/step - loss: 2.2485 - mean_iou: 0.2831 - val_loss: 2.4987 - val_mean_iou: 0.2400 Epoch 30/50 297/297 [==============================] - 249s 835ms/step - loss: 2.2441 - mean_iou: 0.2849 - val_loss: 2.4595 - val_mean_iou: 0.2436 Epoch 31/50 297/297 [==============================] - 249s 835ms/step - loss: 2.2363 - mean_iou: 0.2863 - val_loss: 2.4592 - val_mean_iou: 0.2454 Epoch 32/50 297/297 [==============================] - 249s 835ms/step - loss: 2.2319 - mean_iou: 0.2865 - val_loss: 2.4548 - val_mean_iou: 0.2480 Epoch 33/50 297/297 [==============================] - 245s 821ms/step - loss: 2.2259 - mean_iou: 0.2887 - val_loss: 2.4571 - val_mean_iou: 0.2464 Epoch 34/50 297/297 [==============================] - 248s 830ms/step - loss: 2.2207 - mean_iou: 0.2892 - val_loss: 2.4473 - val_mean_iou: 0.2511 Epoch 35/50 297/297 [==============================] - 246s 825ms/step - loss: 2.2179 - mean_iou: 0.2899 - val_loss: 2.4675 - val_mean_iou: 0.2459 Epoch 36/50 297/297 [==============================] - 244s 816ms/step - loss: 2.2182 - mean_iou: 0.2898 - val_loss: 2.4751 - val_mean_iou: 0.2420 Epoch 37/50 297/297 [==============================] - 249s 835ms/step - loss: 2.2073 - mean_iou: 0.2922 - val_loss: 2.4582 - val_mean_iou: 0.2446 Epoch 38/50 297/297 [==============================] - 251s 841ms/step - loss: 2.2025 - mean_iou: 0.2932 - val_loss: 2.4918 - val_mean_iou: 0.2411 Epoch 39/50 297/297 [==============================] - 246s 824ms/step - loss: 2.1997 - mean_iou: 0.2937 - val_loss: 2.4809 - val_mean_iou: 0.2415 Epoch 40/50 297/297 [==============================] - 245s 820ms/step - loss: 2.1951 - mean_iou: 0.2947 - val_loss: 2.4631 - val_mean_iou: 0.2501 Epoch 41/50 297/297 [==============================] - 249s 833ms/step - loss: 2.1928 - mean_iou: 0.2955 - val_loss: 2.4816 - val_mean_iou: 0.2452 Epoch 42/50 297/297 [==============================] - 248s 830ms/step - loss: 2.1891 - mean_iou: 0.2955 - val_loss: 2.5053 - val_mean_iou: 0.2375 Epoch 43/50 297/297 [==============================] - 247s 827ms/step - loss: 2.1896 - mean_iou: 0.2958 - val_loss: 2.4851 - val_mean_iou: 0.2468 Epoch 44/50 297/297 [==============================] - 245s 822ms/step - loss: 2.1799 - mean_iou: 0.2972 - val_loss: 2.4719 - val_mean_iou: 0.2479 Epoch 45/50 297/297 [==============================] - 247s 828ms/step - loss: 2.1797 - mean_iou: 0.2975 - val_loss: 2.4740 - val_mean_iou: 0.2462 Epoch 46/50 297/297 [==============================] - 247s 827ms/step - loss: 2.1724 - mean_iou: 0.2996 - val_loss: 2.4587 - val_mean_iou: 0.2464 Epoch 47/50 297/297 [==============================] - 247s 827ms/step - loss: 2.1725 - mean_iou: 0.2992 - val_loss: 2.5075 - val_mean_iou: 0.2391 Epoch 48/50 297/297 [==============================] - 247s 826ms/step - loss: 2.1652 - mean_iou: 0.3016 - val_loss: 2.4816 - val_mean_iou: 0.2453 Epoch 00048: early stopping INFO:tensorflow:Assets written to: saved/models/VGG16_U-Net_augmented/assets
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "combine_loss", training_time, show_pred=True)
25/25 [==============================] - 33s 1s/step - loss: 2.2817 - mean_iou: 0.3168 Pour le meilleur modèle on obtient : mean_iou : 0.31677231192588806 loss : 2.2816824913024902 0
1
2
3
4
5
6
7
8
9
# Build model
model = models.build_vgg16_unet(INPUT_SHAPE, NB_CLASSES)
model_name = model.name + '_freeze'
models.freeze_layer_vgg16(model)
freeze layer: block1_conv1 freeze layer: block1_conv2 freeze layer: block1_pool freeze layer: block2_conv1 freeze layer: block2_conv2 freeze layer: block2_pool freeze layer: block3_conv1 freeze layer: block3_conv2 freeze layer: block3_conv3 freeze layer: block3_pool freeze layer: block4_conv1 freeze layer: block4_conv2 freeze layer: block4_conv3 freeze layer: block4_pool
# Train model
print(f'Modèle : {model_name}')
model, hist_df, training_time = train_model(model, model_name, total_loss, augmented=False)
# Draw loss function and metric evolution
helpers.draw_history(hist_df)
# Predict from model trained
y_pred_argmax = predict_model(model, model_name, "combine_loss", training_time, show_pred=True)
25/25 [==============================] - 38s 1s/step - loss: 2.2608 - mean_iou: 0.3193 Pour le meilleur modèle on obtient : mean_iou : 0.3192502558231354 loss : 2.260753870010376 0
1
2
3
4
5
6
7
8
9
pd.DataFrame.from_dict(synthese, orient='index')
| loss type | loss score | mean_iou | Training time | Predict time | |
|---|---|---|---|---|---|
| U-Net_base | categorical_crossentropy | 1.045049 | 0.302060 | 6005.426807 | 1.509681 |
| U-Net_dice_loss | dice_loss | 0.376354 | 0.293561 | 11075.384537 | 0.595256 |
| U-Net_combine_loss | combine_loss | 2.482293 | 0.302627 | 8733.732015 | 0.608075 |
| U-Net_augmented | combine_loss | 2.330191 | 0.303681 | 9423.844568 | 0.646507 |
| ResNet50_U-Net_augmented | combine_loss | 2.311371 | 0.313203 | 12474.215919 | 1.753642 |
| VGG16_U-Net_base | combine_loss | 2.267051 | 0.325019 | 8658.989292 | 2.976111 |
| VGG16_U-Net_augmented | combine_loss | 2.298635 | 0.319571 | 10783.304042 | 0.527344 |
| VGG16_U-Net_freeze | combine_loss | 2.260754 | 0.319250 | 9002.626294 | 0.582971 |